fs/xfs/xfs_dquot.c

   1 /*
   2  * Copyright (c) 2000-2003 Silicon Graphics, Inc.
   3  * All Rights Reserved.
   4  *
   5  * This program is free software; you can redistribute it and/or
   6  * modify it under the terms of the GNU General Public License as
   7  * published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope that it would be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write the Free Software Foundation,
  16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  17  */
  18 #include "xfs.h"
  19 #include "xfs_fs.h"
  20 #include "xfs_format.h"
  21 #include "xfs_log_format.h"
  22 #include "xfs_shared.h"
  23 #include "xfs_trans_resv.h"
  24 #include "xfs_bit.h"
  25 #include "xfs_mount.h"
  26 #include "xfs_inode.h"
  27 #include "xfs_bmap.h"
  28 #include "xfs_bmap_util.h"
  29 #include "xfs_alloc.h"
  30 #include "xfs_quota.h"
  31 #include "xfs_error.h"
  32 #include "xfs_trans.h"
  33 #include "xfs_buf_item.h"
  34 #include "xfs_trans_space.h"
  35 #include "xfs_trans_priv.h"
  36 #include "xfs_qm.h"
  37 #include "xfs_cksum.h"
  38 #include "xfs_trace.h"
  39 #include "xfs_log.h"
  40 #include "xfs_bmap_btree.h"
  41
  42 /*
  43  * Lock order:
  44  *
  45  * ip->i_lock
  46  *   qi->qi_tree_lock
  47  *     dquot->q_qlock (xfs_dqlock() and friends)
  48  *       dquot->q_flush (xfs_dqflock() and friends)
  49  *       qi->qi_lru_lock
  50  *
  51  * If two dquots need to be locked the order is user before group/project,
  52  * otherwise by the lowest id first, see xfs_dqlock2.
  53  */
  54
  55 #ifdef DEBUG
  56 xfs_buftarg_t *xfs_dqerror_target;
  57 int xfs_do_dqerror;
  58 int xfs_dqreq_num;
  59 int xfs_dqerror_mod = 33;
  60 #endif
  61
  62 struct kmem_zone                *xfs_qm_dqtrxzone;
  63 static struct kmem_zone         *xfs_qm_dqzone;
  64
  65 static struct lock_class_key xfs_dquot_group_class;
  66 static struct lock_class_key xfs_dquot_project_class;
  67
  68 /*
  69  * This is called to free all the memory associated with a dquot
  70  */
  71 void
  72 xfs_qm_dqdestroy(
  73         xfs_dquot_t     *dqp)
  74 {
  75         ASSERT(list_empty(&dqp->q_lru));
  76
  77         mutex_destroy(&dqp->q_qlock);
  78
  79         XFS_STATS_DEC(dqp->q_mount, xs_qm_dquot);
  80         kmem_zone_free(xfs_qm_dqzone, dqp);
  81 }
  82
  83 /*
  84  * If default limits are in force, push them into the dquot now.
  85  * We overwrite the dquot limits only if they are zero and this
  86  * is not the root dquot.
  87  */
  88 void
  89 xfs_qm_adjust_dqlimits(
  90         struct xfs_mount        *mp,
  91         struct xfs_dquot        *dq)
  92 {
  93         struct xfs_quotainfo    *q = mp->m_quotainfo;
  94         struct xfs_disk_dquot   *d = &dq->q_core;
  95         int                     prealloc = 0;
  96
  97         ASSERT(d->d_id);
  98
  99         if (q->qi_bsoftlimit && !d->d_blk_softlimit) {
 100                 d->d_blk_softlimit = cpu_to_be64(q->qi_bsoftlimit);
 101                 prealloc = 1;
 102         }
 103         if (q->qi_bhardlimit && !d->d_blk_hardlimit) {
 104                 d->d_blk_hardlimit = cpu_to_be64(q->qi_bhardlimit);
 105                 prealloc = 1;
 106         }
 107         if (q->qi_isoftlimit && !d->d_ino_softlimit)
 108                 d->d_ino_softlimit = cpu_to_be64(q->qi_isoftlimit);
 109         if (q->qi_ihardlimit && !d->d_ino_hardlimit)
 110                 d->d_ino_hardlimit = cpu_to_be64(q->qi_ihardlimit);
 111         if (q->qi_rtbsoftlimit && !d->d_rtb_softlimit)
 112                 d->d_rtb_softlimit = cpu_to_be64(q->qi_rtbsoftlimit);
 113         if (q->qi_rtbhardlimit && !d->d_rtb_hardlimit)
 114                 d->d_rtb_hardlimit = cpu_to_be64(q->qi_rtbhardlimit);
 115
 116         if (prealloc)
 117                 xfs_dquot_set_prealloc_limits(dq);
 118 }
 119
 120 /*
 121  * Check the limits and timers of a dquot and start or reset timers
 122  * if necessary.
 123  * This gets called even when quota enforcement is OFF, which makes our
 124  * life a little less complicated. (We just don't reject any quota
 125  * reservations in that case, when enforcement is off).
 126  * We also return 0 as the values of the timers in Q_GETQUOTA calls, when
 127  * enforcement's off.
 128  * In contrast, warnings are a little different in that they don't
 129  * 'automatically' get started when limits get exceeded.  They do
 130  * get reset to zero, however, when we find the count to be under
 131  * the soft limit (they are only ever set non-zero via userspace).
 132  */
 133 void
 134 xfs_qm_adjust_dqtimers(
 135         xfs_mount_t             *mp,
 136         xfs_disk_dquot_t        *d)
 137 {
 138         ASSERT(d->d_id);
 139
 140 #ifdef DEBUG
 141         if (d->d_blk_hardlimit)
 142                 ASSERT(be64_to_cpu(d->d_blk_softlimit) <=
 143                        be64_to_cpu(d->d_blk_hardlimit));
 144         if (d->d_ino_hardlimit)
 145                 ASSERT(be64_to_cpu(d->d_ino_softlimit) <=
 146                        be64_to_cpu(d->d_ino_hardlimit));
 147         if (d->d_rtb_hardlimit)
 148                 ASSERT(be64_to_cpu(d->d_rtb_softlimit) <=
 149                        be64_to_cpu(d->d_rtb_hardlimit));
 150 #endif
 151
 152         if (!d->d_btimer) {
 153                 if ((d->d_blk_softlimit &&
 154                      (be64_to_cpu(d->d_bcount) >
 155                       be64_to_cpu(d->d_blk_softlimit))) ||
 156                     (d->d_blk_hardlimit &&
 157                      (be64_to_cpu(d->d_bcount) >
 158                       be64_to_cpu(d->d_blk_hardlimit)))) {
 159                         d->d_btimer = cpu_to_be32(get_seconds() +
 160                                         mp->m_quotainfo->qi_btimelimit);
 161                 } else {
 162                         d->d_bwarns = 0;
 163                 }
 164         } else {
 165                 if ((!d->d_blk_softlimit ||
 166                      (be64_to_cpu(d->d_bcount) <=
 167                       be64_to_cpu(d->d_blk_softlimit))) &&
 168                     (!d->d_blk_hardlimit ||
 169                     (be64_to_cpu(d->d_bcount) <=
 170                      be64_to_cpu(d->d_blk_hardlimit)))) {
 171                         d->d_btimer = 0;
 172                 }
 173         }
 174
 175         if (!d->d_itimer) {
 176                 if ((d->d_ino_softlimit &&
 177                      (be64_to_cpu(d->d_icount) >
 178                       be64_to_cpu(d->d_ino_softlimit))) ||
 179                     (d->d_ino_hardlimit &&
 180                      (be64_to_cpu(d->d_icount) >
 181                       be64_to_cpu(d->d_ino_hardlimit)))) {
 182                         d->d_itimer = cpu_to_be32(get_seconds() +
 183                                         mp->m_quotainfo->qi_itimelimit);
 184                 } else {
 185                         d->d_iwarns = 0;
 186                 }
 187         } else {
 188                 if ((!d->d_ino_softlimit ||
 189                      (be64_to_cpu(d->d_icount) <=
 190                       be64_to_cpu(d->d_ino_softlimit)))  &&
 191                     (!d->d_ino_hardlimit ||
 192                      (be64_to_cpu(d->d_icount) <=
 193                       be64_to_cpu(d->d_ino_hardlimit)))) {
 194                         d->d_itimer = 0;
 195                 }
 196         }
 197
 198         if (!d->d_rtbtimer) {
 199                 if ((d->d_rtb_softlimit &&
 200                      (be64_to_cpu(d->d_rtbcount) >
 201                       be64_to_cpu(d->d_rtb_softlimit))) ||
 202                     (d->d_rtb_hardlimit &&
 203                      (be64_to_cpu(d->d_rtbcount) >
 204                       be64_to_cpu(d->d_rtb_hardlimit)))) {
 205                         d->d_rtbtimer = cpu_to_be32(get_seconds() +
 206                                         mp->m_quotainfo->qi_rtbtimelimit);
 207                 } else {
 208                         d->d_rtbwarns = 0;
 209                 }
 210         } else {
 211                 if ((!d->d_rtb_softlimit ||
 212                      (be64_to_cpu(d->d_rtbcount) <=
 213                       be64_to_cpu(d->d_rtb_softlimit))) &&
 214                     (!d->d_rtb_hardlimit ||
 215                      (be64_to_cpu(d->d_rtbcount) <=
 216                       be64_to_cpu(d->d_rtb_hardlimit)))) {
 217                         d->d_rtbtimer = 0;
 218                 }
 219         }
 220 }
 221
 222 /*
 223  * initialize a buffer full of dquots and log the whole thing
 224  */
 225 STATIC void
 226 xfs_qm_init_dquot_blk(
 227         xfs_trans_t     *tp,
 228         xfs_mount_t     *mp,
 229         xfs_dqid_t      id,
 230         uint            type,
 231         xfs_buf_t       *bp)
 232 {
 233         struct xfs_quotainfo    *q = mp->m_quotainfo;
 234         xfs_dqblk_t     *d;
 235         int             curid, i;
 236
 237         ASSERT(tp);
 238         ASSERT(xfs_buf_islocked(bp));
 239
 240         d = bp->b_addr;
 241
 242         /*
 243          * ID of the first dquot in the block - id's are zero based.
 244          */
 245         curid = id - (id % q->qi_dqperchunk);
 246         ASSERT(curid >= 0);
 247         memset(d, 0, BBTOB(q->qi_dqchunklen));
 248         for (i = 0; i < q->qi_dqperchunk; i++, d++, curid++) {
 249                 d->dd_diskdq.d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
 250                 d->dd_diskdq.d_version = XFS_DQUOT_VERSION;
 251                 d->dd_diskdq.d_id = cpu_to_be32(curid);
 252                 d->dd_diskdq.d_flags = type;
 253                 if (xfs_sb_version_hascrc(&mp->m_sb)) {
 254                         uuid_copy(&d->dd_uuid, &mp->m_sb.sb_meta_uuid);
 255                         xfs_update_cksum((char *)d, sizeof(struct xfs_dqblk),
 256                                          XFS_DQUOT_CRC_OFF);
 257                 }
 258         }
 259
 260         xfs_trans_dquot_buf(tp, bp,
 261                             (type & XFS_DQ_USER ? XFS_BLF_UDQUOT_BUF :
 262                             ((type & XFS_DQ_PROJ) ? XFS_BLF_PDQUOT_BUF :
 263                              XFS_BLF_GDQUOT_BUF)));
 264         xfs_trans_log_buf(tp, bp, 0, BBTOB(q->qi_dqchunklen) - 1);
 265 }
 266
 267 /*
 268  * Initialize the dynamic speculative preallocation thresholds. The lo/hi
 269  * watermarks correspond to the soft and hard limits by default. If a soft limit
 270  * is not specified, we use 95% of the hard limit.
 271  */
 272 void
 273 xfs_dquot_set_prealloc_limits(struct xfs_dquot *dqp)
 274 {
 275         __uint64_t space;
 276
 277         dqp->q_prealloc_hi_wmark = be64_to_cpu(dqp->q_core.d_blk_hardlimit);
 278         dqp->q_prealloc_lo_wmark = be64_to_cpu(dqp->q_core.d_blk_softlimit);
 279         if (!dqp->q_prealloc_lo_wmark) {
 280                 dqp->q_prealloc_lo_wmark = dqp->q_prealloc_hi_wmark;
 281                 do_div(dqp->q_prealloc_lo_wmark, 100);
 282                 dqp->q_prealloc_lo_wmark *= 95;
 283         }
 284
 285         space = dqp->q_prealloc_hi_wmark;
 286
 287         do_div(space, 100);
 288         dqp->q_low_space[XFS_QLOWSP_1_PCNT] = space;
 289         dqp->q_low_space[XFS_QLOWSP_3_PCNT] = space * 3;
 290         dqp->q_low_space[XFS_QLOWSP_5_PCNT] = space * 5;
 291 }
 292
 293 /*
 294  * Allocate a block and fill it with dquots.
 295  * This is called when the bmapi finds a hole.
 296  */
 297 STATIC int
 298 xfs_qm_dqalloc(
 299         xfs_trans_t     **tpp,
 300         xfs_mount_t     *mp,
 301         xfs_dquot_t     *dqp,
 302         xfs_inode_t     *quotip,
 303         xfs_fileoff_t   offset_fsb,
 304         xfs_buf_t       **O_bpp)
 305 {
 306         xfs_fsblock_t   firstblock;
 307         xfs_bmap_free_t flist;
 308         xfs_bmbt_irec_t map;
 309         int             nmaps, error;
 310         xfs_buf_t       *bp;
 311         xfs_trans_t     *tp = *tpp;
 312
 313         ASSERT(tp != NULL);
 314
 315         trace_xfs_dqalloc(dqp);
 316
 317         /*
 318          * Initialize the bmap freelist prior to calling bmapi code.
 319          */
 320         xfs_bmap_init(&flist, &firstblock);
 321         xfs_ilock(quotip, XFS_ILOCK_EXCL);
 322         /*
 323          * Return if this type of quotas is turned off while we didn't
 324          * have an inode lock
 325          */
 326         if (!xfs_this_quota_on(dqp->q_mount, dqp->dq_flags)) {
 327                 xfs_iunlock(quotip, XFS_ILOCK_EXCL);
 328                 return -ESRCH;
 329         }
 330
 331         xfs_trans_ijoin(tp, quotip, XFS_ILOCK_EXCL);
 332         nmaps = 1;
 333         error = xfs_bmapi_write(tp, quotip, offset_fsb,
 334                                 XFS_DQUOT_CLUSTER_SIZE_FSB, XFS_BMAPI_METADATA,
 335                                 &firstblock, XFS_QM_DQALLOC_SPACE_RES(mp),
 336                                 &map, &nmaps, &flist);
 337         if (error)
 338                 goto error0;
 339         ASSERT(map.br_blockcount == XFS_DQUOT_CLUSTER_SIZE_FSB);
 340         ASSERT(nmaps == 1);
 341         ASSERT((map.br_startblock != DELAYSTARTBLOCK) &&
 342                (map.br_startblock != HOLESTARTBLOCK));
 343
 344         /*
 345          * Keep track of the blkno to save a lookup later
 346          */
 347         dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
 348
 349         /* now we can just get the buffer (there's nothing to read yet) */
 350         bp = xfs_trans_get_buf(tp, mp->m_ddev_targp,
 351                                dqp->q_blkno,
 352                                mp->m_quotainfo->qi_dqchunklen,
 353                                0);
 354         if (!bp) {
 355                 error = -ENOMEM;
 356                 goto error1;
 357         }
 358         bp->b_ops = &xfs_dquot_buf_ops;
 359
 360         /*
 361          * Make a chunk of dquots out of this buffer and log
 362          * the entire thing.
 363          */
 364         xfs_qm_init_dquot_blk(tp, mp, be32_to_cpu(dqp->q_core.d_id),
 365                               dqp->dq_flags & XFS_DQ_ALLTYPES, bp);
 366
 367         /*
 368          * xfs_bmap_finish() may commit the current transaction and
 369          * start a second transaction if the freelist is not empty.
 370          *
 371          * Since we still want to modify this buffer, we need to
 372          * ensure that the buffer is not released on commit of
 373          * the first transaction and ensure the buffer is added to the
 374          * second transaction.
 375          *
 376          * If there is only one transaction then don't stop the buffer
 377          * from being released when it commits later on.
 378          */
 379
 380         xfs_trans_bhold(tp, bp);
 381
 382         error = xfs_bmap_finish(tpp, &flist, NULL);
 383         if (error)
 384                 goto error1;
 385
 386         /* Transaction was committed? */
 387         if (*tpp != tp) {
 388                 tp = *tpp;
 389                 xfs_trans_bjoin(tp, bp);
 390         } else {
 391                 xfs_trans_bhold_release(tp, bp);
 392         }
 393
 394         *O_bpp = bp;
 395         return 0;
 396
 397 error1:
 398         xfs_bmap_cancel(&flist);
 399 error0:
 400         xfs_iunlock(quotip, XFS_ILOCK_EXCL);
 401
 402         return error;
 403 }
 404
 405 STATIC int
 406 xfs_qm_dqrepair(
 407         struct xfs_mount        *mp,
 408         struct xfs_trans        *tp,
 409         struct xfs_dquot        *dqp,
 410         xfs_dqid_t              firstid,
 411         struct xfs_buf          **bpp)
 412 {
 413         int                     error;
 414         struct xfs_disk_dquot   *ddq;
 415         struct xfs_dqblk        *d;
 416         int                     i;
 417
 418         /*
 419          * Read the buffer without verification so we get the corrupted
 420          * buffer returned to us. make sure we verify it on write, though.
 421          */
 422         error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp, dqp->q_blkno,
 423                                    mp->m_quotainfo->qi_dqchunklen,
 424                                    0, bpp, NULL);
 425
 426         if (error) {
 427                 ASSERT(*bpp == NULL);
 428                 return error;
 429         }
 430         (*bpp)->b_ops = &xfs_dquot_buf_ops;
 431
 432         ASSERT(xfs_buf_islocked(*bpp));
 433         d = (struct xfs_dqblk *)(*bpp)->b_addr;
 434
 435         /* Do the actual repair of dquots in this buffer */
 436         for (i = 0; i < mp->m_quotainfo->qi_dqperchunk; i++) {
 437                 ddq = &d[i].dd_diskdq;
 438                 error = xfs_dqcheck(mp, ddq, firstid + i,
 439                                        dqp->dq_flags & XFS_DQ_ALLTYPES,
 440                                        XFS_QMOPT_DQREPAIR, "xfs_qm_dqrepair");
 441                 if (error) {
 442                         /* repair failed, we're screwed */
 443                         xfs_trans_brelse(tp, *bpp);
 444                         return -EIO;
 445                 }
 446         }
 447
 448         return 0;
 449 }
 450
 451 /*
 452  * Maps a dquot to the buffer containing its on-disk version.
 453  * This returns a ptr to the buffer containing the on-disk dquot
 454  * in the bpp param, and a ptr to the on-disk dquot within that buffer
 455  */
 456 STATIC int
 457 xfs_qm_dqtobp(
 458         xfs_trans_t             **tpp,
 459         xfs_dquot_t             *dqp,
 460         xfs_disk_dquot_t        **O_ddpp,
 461         xfs_buf_t               **O_bpp,
 462         uint                    flags)
 463 {
 464         struct xfs_bmbt_irec    map;
 465         int                     nmaps = 1, error;
 466         struct xfs_buf          *bp;
 467         struct xfs_inode        *quotip = xfs_dq_to_quota_inode(dqp);
 468         struct xfs_mount        *mp = dqp->q_mount;
 469         xfs_dqid_t              id = be32_to_cpu(dqp->q_core.d_id);
 470         struct xfs_trans        *tp = (tpp ? *tpp : NULL);
 471         uint                    lock_mode;
 472
 473         dqp->q_fileoffset = (xfs_fileoff_t)id / mp->m_quotainfo->qi_dqperchunk;
 474
 475         lock_mode = xfs_ilock_data_map_shared(quotip);
 476         if (!xfs_this_quota_on(dqp->q_mount, dqp->dq_flags)) {
 477                 /*
 478                  * Return if this type of quotas is turned off while we
 479                  * didn't have the quota inode lock.
 480                  */
 481                 xfs_iunlock(quotip, lock_mode);
 482                 return -ESRCH;
 483         }
 484
 485         /*
 486          * Find the block map; no allocations yet
 487          */
 488         error = xfs_bmapi_read(quotip, dqp->q_fileoffset,
 489                                XFS_DQUOT_CLUSTER_SIZE_FSB, &map, &nmaps, 0);
 490
 491         xfs_iunlock(quotip, lock_mode);
 492         if (error)
 493                 return error;
 494
 495         ASSERT(nmaps == 1);
 496         ASSERT(map.br_blockcount == 1);
 497
 498         /*
 499          * Offset of dquot in the (fixed sized) dquot chunk.
 500          */
 501         dqp->q_bufoffset = (id % mp->m_quotainfo->qi_dqperchunk) *
 502                 sizeof(xfs_dqblk_t);
 503
 504         ASSERT(map.br_startblock != DELAYSTARTBLOCK);
 505         if (map.br_startblock == HOLESTARTBLOCK) {
 506                 /*
 507                  * We don't allocate unless we're asked to
 508                  */
 509                 if (!(flags & XFS_QMOPT_DQALLOC))
 510                         return -ENOENT;
 511
 512                 ASSERT(tp);
 513                 error = xfs_qm_dqalloc(tpp, mp, dqp, quotip,
 514                                         dqp->q_fileoffset, &bp);
 515                 if (error)
 516                         return error;
 517                 tp = *tpp;
 518         } else {
 519                 trace_xfs_dqtobp_read(dqp);
 520
 521                 /*
 522                  * store the blkno etc so that we don't have to do the
 523                  * mapping all the time
 524                  */
 525                 dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
 526
 527                 error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
 528                                            dqp->q_blkno,
 529                                            mp->m_quotainfo->qi_dqchunklen,
 530                                            0, &bp, &xfs_dquot_buf_ops);
 531
 532                 if (error == -EFSCORRUPTED && (flags & XFS_QMOPT_DQREPAIR)) {
 533                         xfs_dqid_t firstid = (xfs_dqid_t)map.br_startoff *
 534                                                 mp->m_quotainfo->qi_dqperchunk;
 535                         ASSERT(bp == NULL);
 536                         error = xfs_qm_dqrepair(mp, tp, dqp, firstid, &bp);
 537                 }
 538
 539                 if (error) {
 540                         ASSERT(bp == NULL);
 541                         return error;
 542                 }
 543         }
 544
 545         ASSERT(xfs_buf_islocked(bp));
 546         *O_bpp = bp;
 547         *O_ddpp = bp->b_addr + dqp->q_bufoffset;
 548
 549         return 0;
 550 }
 551
 552
 553 /*
 554  * Read in the ondisk dquot using dqtobp() then copy it to an incore version,
 555  * and release the buffer immediately.
 556  *
 557  * If XFS_QMOPT_DQALLOC is set, allocate a dquot on disk if it needed.
 558  */
 559 int
 560 xfs_qm_dqread(
 561         struct xfs_mount        *mp,
 562         xfs_dqid_t              id,
 563         uint                    type,
 564         uint                    flags,
 565         struct xfs_dquot        **O_dqpp)
 566 {
 567         struct xfs_dquot        *dqp;
 568         struct xfs_disk_dquot   *ddqp;
 569         struct xfs_buf          *bp;
 570         struct xfs_trans        *tp = NULL;
 571         int                     error;
 572
 573         dqp = kmem_zone_zalloc(xfs_qm_dqzone, KM_SLEEP);
 574
 575         dqp->dq_flags = type;
 576         dqp->q_core.d_id = cpu_to_be32(id);
 577         dqp->q_mount = mp;
 578         INIT_LIST_HEAD(&dqp->q_lru);
 579         mutex_init(&dqp->q_qlock);
 580         init_waitqueue_head(&dqp->q_pinwait);
 581
 582         /*
 583          * Because we want to use a counting completion, complete
 584          * the flush completion once to allow a single access to
 585          * the flush completion without blocking.
 586          */
 587         init_completion(&dqp->q_flush);
 588         complete(&dqp->q_flush);
 589
 590         /*
 591          * Make sure group quotas have a different lock class than user
 592          * quotas.
 593          */
 594         switch (type) {
 595         case XFS_DQ_USER:
 596                 /* uses the default lock class */
 597                 break;
 598         case XFS_DQ_GROUP:
 599                 lockdep_set_class(&dqp->q_qlock, &xfs_dquot_group_class);
 600                 break;
 601         case XFS_DQ_PROJ:
 602                 lockdep_set_class(&dqp->q_qlock, &xfs_dquot_project_class);
 603                 break;
 604         default:
 605                 ASSERT(0);
 606                 break;
 607         }
 608
 609         XFS_STATS_INC(mp, xs_qm_dquot);
 610
 611         trace_xfs_dqread(dqp);
 612
 613         if (flags & XFS_QMOPT_DQALLOC) {
 614                 tp = xfs_trans_alloc(mp, XFS_TRANS_QM_DQALLOC);
 615                 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_qm_dqalloc,
 616                                           XFS_QM_DQALLOC_SPACE_RES(mp), 0);
 617                 if (error)
 618                         goto error1;
 619         }
 620
 621         /*
 622          * get a pointer to the on-disk dquot and the buffer containing it
 623          * dqp already knows its own type (GROUP/USER).
 624          */
 625         error = xfs_qm_dqtobp(&tp, dqp, &ddqp, &bp, flags);
 626         if (error) {
 627                 /*
 628                  * This can happen if quotas got turned off (ESRCH),
 629                  * or if the dquot didn't exist on disk and we ask to
 630                  * allocate (ENOENT).
 631                  */
 632                 trace_xfs_dqread_fail(dqp);
 633                 goto error1;
 634         }
 635
 636         /* copy everything from disk dquot to the incore dquot */
 637         memcpy(&dqp->q_core, ddqp, sizeof(xfs_disk_dquot_t));
 638         xfs_qm_dquot_logitem_init(dqp);
 639
 640         /*
 641          * Reservation counters are defined as reservation plus current usage
 642          * to avoid having to add every time.
 643          */
 644         dqp->q_res_bcount = be64_to_cpu(ddqp->d_bcount);
 645         dqp->q_res_icount = be64_to_cpu(ddqp->d_icount);
 646         dqp->q_res_rtbcount = be64_to_cpu(ddqp->d_rtbcount);
 647
 648         /* initialize the dquot speculative prealloc thresholds */
 649         xfs_dquot_set_prealloc_limits(dqp);
 650
 651         /* Mark the buf so that this will stay incore a little longer */
 652         xfs_buf_set_ref(bp, XFS_DQUOT_REF);
 653
 654         /*
 655          * We got the buffer with a xfs_trans_read_buf() (in dqtobp())
 656          * So we need to release with xfs_trans_brelse().
 657          * The strategy here is identical to that of inodes; we lock
 658          * the dquot in xfs_qm_dqget() before making it accessible to
 659          * others. This is because dquots, like inodes, need a good level of
 660          * concurrency, and we don't want to take locks on the entire buffers
 661          * for dquot accesses.
 662          * Note also that the dquot buffer may even be dirty at this point, if
 663          * this particular dquot was repaired. We still aren't afraid to
 664          * brelse it because we have the changes incore.
 665          */
 666         ASSERT(xfs_buf_islocked(bp));
 667         xfs_trans_brelse(tp, bp);
 668
 669         if (tp) {
 670                 error = xfs_trans_commit(tp);
 671                 if (error)
 672                         goto error0;
 673         }
 674
 675         *O_dqpp = dqp;
 676         return error;
 677
 678 error1:
 679         if (tp)
 680                 xfs_trans_cancel(tp);
 681 error0:
 682         xfs_qm_dqdestroy(dqp);
 683         *O_dqpp = NULL;
 684         return error;
 685 }
 686
 687 /*
 688  * Given the file system, inode OR id, and type (UDQUOT/GDQUOT), return a
 689  * a locked dquot, doing an allocation (if requested) as needed.
 690  * When both an inode and an id are given, the inode's id takes precedence.
 691  * That is, if the id changes while we don't hold the ilock inside this
 692  * function, the new dquot is returned, not necessarily the one requested
 693  * in the id argument.
 694  */
 695 int
 696 xfs_qm_dqget(
 697         xfs_mount_t     *mp,
 698         xfs_inode_t     *ip,      /* locked inode (optional) */
 699         xfs_dqid_t      id,       /* uid/projid/gid depending on type */
 700         uint            type,     /* XFS_DQ_USER/XFS_DQ_PROJ/XFS_DQ_GROUP */
 701         uint            flags,    /* DQALLOC, DQSUSER, DQREPAIR, DOWARN */
 702         xfs_dquot_t     **O_dqpp) /* OUT : locked incore dquot */
 703 {
 704         struct xfs_quotainfo    *qi = mp->m_quotainfo;
 705         struct radix_tree_root *tree = xfs_dquot_tree(qi, type);
 706         struct xfs_dquot        *dqp;
 707         int                     error;
 708
 709         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
 710         if ((! XFS_IS_UQUOTA_ON(mp) && type == XFS_DQ_USER) ||
 711             (! XFS_IS_PQUOTA_ON(mp) && type == XFS_DQ_PROJ) ||
 712             (! XFS_IS_GQUOTA_ON(mp) && type == XFS_DQ_GROUP)) {
 713                 return -ESRCH;
 714         }
 715
 716 #ifdef DEBUG
 717         if (xfs_do_dqerror) {
 718                 if ((xfs_dqerror_target == mp->m_ddev_targp) &&
 719                     (xfs_dqreq_num++ % xfs_dqerror_mod) == 0) {
 720                         xfs_debug(mp, "Returning error in dqget");
 721                         return -EIO;
 722                 }
 723         }
 724
 725         ASSERT(type == XFS_DQ_USER ||
 726                type == XFS_DQ_PROJ ||
 727                type == XFS_DQ_GROUP);
 728         if (ip) {
 729                 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 730                 ASSERT(xfs_inode_dquot(ip, type) == NULL);
 731         }
 732 #endif
 733
 734 restart:
 735         mutex_lock(&qi->qi_tree_lock);
 736         dqp = radix_tree_lookup(tree, id);
 737         if (dqp) {
 738                 xfs_dqlock(dqp);
 739                 if (dqp->dq_flags & XFS_DQ_FREEING) {
 740                         xfs_dqunlock(dqp);
 741                         mutex_unlock(&qi->qi_tree_lock);
 742                         trace_xfs_dqget_freeing(dqp);
 743                         delay(1);
 744                         goto restart;
 745                 }
 746
 747                 dqp->q_nrefs++;
 748                 mutex_unlock(&qi->qi_tree_lock);
 749
 750                 trace_xfs_dqget_hit(dqp);
 751                 XFS_STATS_INC(mp, xs_qm_dqcachehits);
 752                 *O_dqpp = dqp;
 753                 return 0;
 754         }
 755         mutex_unlock(&qi->qi_tree_lock);
 756         XFS_STATS_INC(mp, xs_qm_dqcachemisses);
 757
 758         /*
 759          * Dquot cache miss. We don't want to keep the inode lock across
 760          * a (potential) disk read. Also we don't want to deal with the lock
 761          * ordering between quotainode and this inode. OTOH, dropping the inode
 762          * lock here means dealing with a chown that can happen before
 763          * we re-acquire the lock.
 764          */
 765         if (ip)
 766                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 767
 768         error = xfs_qm_dqread(mp, id, type, flags, &dqp);
 769
 770         if (ip)
 771                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 772
 773         if (error)
 774                 return error;
 775
 776         if (ip) {
 777                 /*
 778                  * A dquot could be attached to this inode by now, since
 779                  * we had dropped the ilock.
 780                  */
 781                 if (xfs_this_quota_on(mp, type)) {
 782                         struct xfs_dquot        *dqp1;
 783
 784                         dqp1 = xfs_inode_dquot(ip, type);
 785                         if (dqp1) {
 786                                 xfs_qm_dqdestroy(dqp);
 787                                 dqp = dqp1;
 788                                 xfs_dqlock(dqp);
 789                                 goto dqret;
 790                         }
 791                 } else {
 792                         /* inode stays locked on return */
 793                         xfs_qm_dqdestroy(dqp);
 794                         return -ESRCH;
 795                 }
 796         }
 797
 798         mutex_lock(&qi->qi_tree_lock);
 799         error = radix_tree_insert(tree, id, dqp);
 800         if (unlikely(error)) {
 801                 WARN_ON(error != -EEXIST);
 802
 803                 /*
 804                  * Duplicate found. Just throw away the new dquot and start
 805                  * over.
 806                  */
 807                 mutex_unlock(&qi->qi_tree_lock);
 808                 trace_xfs_dqget_dup(dqp);
 809                 xfs_qm_dqdestroy(dqp);
 810                 XFS_STATS_INC(mp, xs_qm_dquot_dups);
 811                 goto restart;
 812         }
 813
 814         /*
 815          * We return a locked dquot to the caller, with a reference taken
 816          */
 817         xfs_dqlock(dqp);
 818         dqp->q_nrefs = 1;
 819
 820         qi->qi_dquots++;
 821         mutex_unlock(&qi->qi_tree_lock);
 822
 823  dqret:
 824         ASSERT((ip == NULL) || xfs_isilocked(ip, XFS_ILOCK_EXCL));
 825         trace_xfs_dqget_miss(dqp);
 826         *O_dqpp = dqp;
 827         return 0;
 828 }
 829
 830 /*
 831  * Release a reference to the dquot (decrement ref-count) and unlock it.
 832  *
 833  * If there is a group quota attached to this dquot, carefully release that
 834  * too without tripping over deadlocks'n'stuff.
 835  */
 836 void
 837 xfs_qm_dqput(
 838         struct xfs_dquot        *dqp)
 839 {
 840         ASSERT(dqp->q_nrefs > 0);
 841         ASSERT(XFS_DQ_IS_LOCKED(dqp));
 842
 843         trace_xfs_dqput(dqp);
 844
 845         if (--dqp->q_nrefs == 0) {
 846                 struct xfs_quotainfo    *qi = dqp->q_mount->m_quotainfo;
 847                 trace_xfs_dqput_free(dqp);
 848
 849                 if (list_lru_add(&qi->qi_lru, &dqp->q_lru))
 850                         XFS_STATS_INC(dqp->q_mount, xs_qm_dquot_unused);
 851         }
 852         xfs_dqunlock(dqp);
 853 }
 854
 855 /*
 856  * Release a dquot. Flush it if dirty, then dqput() it.
 857  * dquot must not be locked.
 858  */
 859 void
 860 xfs_qm_dqrele(
 861         xfs_dquot_t     *dqp)
 862 {
 863         if (!dqp)
 864                 return;
 865
 866         trace_xfs_dqrele(dqp);
 867
 868         xfs_dqlock(dqp);
 869         /*
 870          * We don't care to flush it if the dquot is dirty here.
 871          * That will create stutters that we want to avoid.
 872          * Instead we do a delayed write when we try to reclaim
 873          * a dirty dquot. Also xfs_sync will take part of the burden...
 874          */
 875         xfs_qm_dqput(dqp);
 876 }
 877
 878 /*
 879  * This is the dquot flushing I/O completion routine.  It is called
 880  * from interrupt level when the buffer containing the dquot is
 881  * flushed to disk.  It is responsible for removing the dquot logitem
 882  * from the AIL if it has not been re-logged, and unlocking the dquot's
 883  * flush lock. This behavior is very similar to that of inodes..
 884  */
 885 STATIC void
 886 xfs_qm_dqflush_done(
 887         struct xfs_buf          *bp,
 888         struct xfs_log_item     *lip)
 889 {
 890         xfs_dq_logitem_t        *qip = (struct xfs_dq_logitem *)lip;
 891         xfs_dquot_t             *dqp = qip->qli_dquot;
 892         struct xfs_ail          *ailp = lip->li_ailp;
 893
 894         /*
 895          * We only want to pull the item from the AIL if its
 896          * location in the log has not changed since we started the flush.
 897          * Thus, we only bother if the dquot's lsn has
 898          * not changed. First we check the lsn outside the lock
 899          * since it's cheaper, and then we recheck while
 900          * holding the lock before removing the dquot from the AIL.
 901          */
 902         if ((lip->li_flags & XFS_LI_IN_AIL) &&
 903             lip->li_lsn == qip->qli_flush_lsn) {
 904
 905                 /* xfs_trans_ail_delete() drops the AIL lock. */
 906                 spin_lock(&ailp->xa_lock);
 907                 if (lip->li_lsn == qip->qli_flush_lsn)
 908                         xfs_trans_ail_delete(ailp, lip, SHUTDOWN_CORRUPT_INCORE);
 909                 else
 910                         spin_unlock(&ailp->xa_lock);
 911         }
 912
 913         /*
 914          * Release the dq's flush lock since we're done with it.
 915          */
 916         xfs_dqfunlock(dqp);
 917 }
 918
 919 /*
 920  * Write a modified dquot to disk.
 921  * The dquot must be locked and the flush lock too taken by caller.
 922  * The flush lock will not be unlocked until the dquot reaches the disk,
 923  * but the dquot is free to be unlocked and modified by the caller
 924  * in the interim. Dquot is still locked on return. This behavior is
 925  * identical to that of inodes.
 926  */
 927 int
 928 xfs_qm_dqflush(
 929         struct xfs_dquot        *dqp,
 930         struct xfs_buf          **bpp)
 931 {
 932         struct xfs_mount        *mp = dqp->q_mount;
 933         struct xfs_buf          *bp;
 934         struct xfs_disk_dquot   *ddqp;
 935         int                     error;
 936
 937         ASSERT(XFS_DQ_IS_LOCKED(dqp));
 938         ASSERT(!completion_done(&dqp->q_flush));
 939
 940         trace_xfs_dqflush(dqp);
 941
 942         *bpp = NULL;
 943
 944         xfs_qm_dqunpin_wait(dqp);
 945
 946         /*
 947          * This may have been unpinned because the filesystem is shutting
 948          * down forcibly. If that's the case we must not write this dquot
 949          * to disk, because the log record didn't make it to disk.
 950          *
 951          * We also have to remove the log item from the AIL in this case,
 952          * as we wait for an emptry AIL as part of the unmount process.
 953          */
 954         if (XFS_FORCED_SHUTDOWN(mp)) {
 955                 struct xfs_log_item     *lip = &dqp->q_logitem.qli_item;
 956                 dqp->dq_flags &= ~XFS_DQ_DIRTY;
 957
 958                 xfs_trans_ail_remove(lip, SHUTDOWN_CORRUPT_INCORE);
 959
 960                 error = -EIO;
 961                 goto out_unlock;
 962         }
 963
 964         /*
 965          * Get the buffer containing the on-disk dquot
 966          */
 967         error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp, dqp->q_blkno,
 968                                    mp->m_quotainfo->qi_dqchunklen, 0, &bp,
 969                                    &xfs_dquot_buf_ops);
 970         if (error)
 971                 goto out_unlock;
 972
 973         /*
 974          * Calculate the location of the dquot inside the buffer.
 975          */
 976         ddqp = bp->b_addr + dqp->q_bufoffset;
 977
 978         /*
 979          * A simple sanity check in case we got a corrupted dquot..
 980          */
 981         error = xfs_dqcheck(mp, &dqp->q_core, be32_to_cpu(ddqp->d_id), 0,
 982                            XFS_QMOPT_DOWARN, "dqflush (incore copy)");
 983         if (error) {
 984                 xfs_buf_relse(bp);
 985                 xfs_dqfunlock(dqp);
 986                 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
 987                 return -EIO;
 988         }
 989
 990         /* This is the only portion of data that needs to persist */
 991         memcpy(ddqp, &dqp->q_core, sizeof(xfs_disk_dquot_t));
 992
 993         /*
 994          * Clear the dirty field and remember the flush lsn for later use.
 995          */
 996         dqp->dq_flags &= ~XFS_DQ_DIRTY;
 997
 998         xfs_trans_ail_copy_lsn(mp->m_ail, &dqp->q_logitem.qli_flush_lsn,
 999                                         &dqp->q_logitem.qli_item.li_lsn);
1000
1001         /*
1002          * copy the lsn into the on-disk dquot now while we have the in memory
1003          * dquot here. This can't be done later in the write verifier as we
1004          * can't get access to the log item at that point in time.
1005          *
1006          * We also calculate the CRC here so that the on-disk dquot in the
1007          * buffer always has a valid CRC. This ensures there is no possibility
1008          * of a dquot without an up-to-date CRC getting to disk.
1009          */
1010         if (xfs_sb_version_hascrc(&mp->m_sb)) {
1011                 struct xfs_dqblk *dqb = (struct xfs_dqblk *)ddqp;
1012
1013                 dqb->dd_lsn = cpu_to_be64(dqp->q_logitem.qli_item.li_lsn);
1014                 xfs_update_cksum((char *)dqb, sizeof(struct xfs_dqblk),
1015                                  XFS_DQUOT_CRC_OFF);
1016         }
1017
1018         /*
1019          * Attach an iodone routine so that we can remove this dquot from the
1020          * AIL and release the flush lock once the dquot is synced to disk.
1021          */
1022         xfs_buf_attach_iodone(bp, xfs_qm_dqflush_done,
1023                                   &dqp->q_logitem.qli_item);
1024
1025         /*
1026          * If the buffer is pinned then push on the log so we won't
1027          * get stuck waiting in the write for too long.
1028          */
1029         if (xfs_buf_ispinned(bp)) {
1030                 trace_xfs_dqflush_force(dqp);
1031                 xfs_log_force(mp, 0);
1032         }
1033
1034         trace_xfs_dqflush_done(dqp);
1035         *bpp = bp;
1036         return 0;
1037
1038 out_unlock:
1039         xfs_dqfunlock(dqp);
1040         return -EIO;
1041 }
1042
1043 /*
1044  * Lock two xfs_dquot structures.
1045  *
1046  * To avoid deadlocks we always lock the quota structure with
1047  * the lowerd id first.
1048  */
1049 void
1050 xfs_dqlock2(
1051         xfs_dquot_t     *d1,
1052         xfs_dquot_t     *d2)
1053 {
1054         if (d1 && d2) {
1055                 ASSERT(d1 != d2);
1056                 if (be32_to_cpu(d1->q_core.d_id) >
1057                     be32_to_cpu(d2->q_core.d_id)) {
1058                         mutex_lock(&d2->q_qlock);
1059                         mutex_lock_nested(&d1->q_qlock, XFS_QLOCK_NESTED);
1060                 } else {
1061                         mutex_lock(&d1->q_qlock);
1062                         mutex_lock_nested(&d2->q_qlock, XFS_QLOCK_NESTED);
1063                 }
1064         } else if (d1) {
1065                 mutex_lock(&d1->q_qlock);
1066         } else if (d2) {
1067                 mutex_lock(&d2->q_qlock);
1068         }
1069 }
1070
1071 int __init
1072 xfs_qm_init(void)
1073 {
1074         xfs_qm_dqzone =
1075                 kmem_zone_init(sizeof(struct xfs_dquot), "xfs_dquot");
1076         if (!xfs_qm_dqzone)
1077                 goto out;
1078
1079         xfs_qm_dqtrxzone =
1080                 kmem_zone_init(sizeof(struct xfs_dquot_acct), "xfs_dqtrx");
1081         if (!xfs_qm_dqtrxzone)
1082                 goto out_free_dqzone;
1083
1084         return 0;
1085
1086 out_free_dqzone:
1087         kmem_zone_destroy(xfs_qm_dqzone);
1088 out:
1089         return -ENOMEM;
1090 }
1091
1092 void
1093 xfs_qm_exit(void)
1094 {
1095         kmem_zone_destroy(xfs_qm_dqtrxzone);
1096         kmem_zone_destroy(xfs_qm_dqzone);
1097 }